The Signal
MASQ preprint reports 8,098-fold QEC sampling speedup
A preprint introduces MASQ, a sampler for near-Clifford quantum error-correction circuits that delays processing some non-Clifford rotations until measurements need them. Across 19 benchmarks, its authors report up to 8,098-fold higher CPU throughput than Clifft on multi-round circuits.

Sampling rare failures
Across 19 quantum error-correction benchmarks, a new sampler reduced peak active width by as many as 40 active virtual qubits and reached up to 8,098-fold higher CPU throughput than Clifft on multi-round circuits, according to a preprint describing the method. The paper introduces MASQ, an exact measurement-adaptive sampler for near-Clifford QEC circuits, and reports that it preserves the measurement-record distribution.
Classical sampling is used to evaluate QEC circuits, but estimating rare logical failures requires many independent shots. The authors identify the dense active-state vector used by frame-active-state simulators as a computational bottleneck: processing costs grow exponentially with the number of active virtual qubits.
The paper, by Chaebin Jung, Sangmin Jeon, Kyeongwon Lee, Sechan Park and Woojoo Lee, was first submitted to arXiv on November 17, 2026. It is a preprint.
Rotations wait for measurements
MASQ targets near-Clifford circuits, which include non-Clifford rotations that cannot be represented by a single stabilizer tableau or Pauli frame. Rather than immediately adding each such rotation to the dense active-state representation, the method leaves it pending until a measurement requires its effects.
At each measurement boundary, MASQ identifies a measurement-relevant closure: the rotations that directly affect that measurement, plus earlier rotations linked to them through non-commuting dependencies. It materializes the rotations in that set in circuit order and defers the others. The authors say this rule preserves the same conditional measurement probabilities as immediate processing of all pending rotations.
A second step, stabilizer-assisted elimination, removes rotations the method certifies as invisible to all future measurements. The approach also computes a circuit-dependent schedule once and reuses it across shots. Adaptive caching stores conditional Born probabilities for recurring measurement contexts; batched execution is another part of the sampling workflow.
Benchmark results and limits
The evaluation covers 19 circuits, including surface-code memories with coherent RZ errors at distances 3, 5, 7 and 9, an RX-error case at distance 3, magic-state distillation and cultivation circuits, and a pure-Clifford surface-code memory. The set also includes multi-round RZ circuits at distances 7 and 9.
On a multi-round surface-code RZ benchmark at distance 5 with three rounds, MASQ reached 7.45 × 10^4 shots per second. The authors report that as 8,098 times Clifft’s CPU throughput and 1,267 times CPU SymFT’s. For magic-state distillation, MASQ reached 3.97 × 10^7 shots per second, which the paper gives as 7.65 times Clifft and 3.68 times CPU SymFT.
Across the full test set, MASQ sampled all 19 circuits and had the highest throughput on 15. Clifft and SymFT each ran out of memory on four circuits, while Tsim exceeded the compilation limit on 12. Those results were measured at a Pauli error rate of 10^-3.
The comparisons were not uniformly in MASQ’s favor: CPU Tsim was fastest on every single-round RZ circuit, with throughput 2.29 to 4.16 times MASQ’s. On a distance-9 case, MASQ completed sampling within a factor of 2.29 of Tsim, while Clifft and SymFT ran out of memory.
Active width and validation
The paper reports peak active-width reductions of 22 to 13 virtual qubits for a distance-5 multi-round RZ circuit and 45 to 25 for a distance-7 circuit. In the distance-5, five-round case, Clifft reached 2^22 amplitudes, while MASQ’s peak was 2^13, a 512-fold reduction in peak dense-state size.
For correctness checks against Clifft, the authors compared matched noise realizations and measurement outcomes across 15 circuits. The maximum absolute Born-probability difference was 1.17 × 10^-17, which the paper describes as within double-precision roundoff.